AraScholar — Academic Paraphrasing in Modern Standard Arabic
AraScholar rewrites Modern Standard Arabic (MSA) sentences into a polished academic register while preserving meaning. It is an AraT5v2-base model fine-tuned on a manually curated corpus of expert academic rewrites.
It accompanies the paper "Reaching the Human Operating Point: Controlled Academic Paraphrasing in Modern Standard Arabic."
The idea: an operating point, not just fluency
Academic rewriting must balance two opposing goals: preserve the source meaning (fidelity) and genuinely rewrite it (novelty). The human gold rewrites sit at a narrow operating point — about 57% novelty at ~87 semantic preservation. Plain decoding of fine-tuned models tends to copy the input; large LLMs tend to over-rewrite and drift. AraScholar is designed to be steered to the human operating point at decoding time.
Usage
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
m = "Omartificial-Intelligence-Space/AraScholar"
tok = AutoTokenizer.from_pretrained(m) # see note below for transformers>=5
model = AutoModelForSeq2SeqLM.from_pretrained(m).eval()
src = "تعد القراءة من اهم وسائل اكتساب المعرفة لدى الطلاب في الجامعات"
ids = tok(f"paraphrase: {src}", return_tensors="pt", truncation=True, max_length=768)
out = model.generate(**ids, num_beams=5, no_repeat_ngram_size=3,
repetition_penalty=1.2, max_new_tokens=512)
print(tok.decode(out[0], skip_special_tokens=True))
Input format: paraphrase: {text} (optionally paraphrase: {domain}: {text}).
Text is used without orthographic normalization.
transformers >= 5 note: if the tokenizer fails to load, use the fast file directly:
from transformers import PreTrainedTokenizerFast; tok = PreTrainedTokenizerFast(tokenizer_file="tokenizer.json", pad_token="<pad>", eos_token="</s>", unk_token="<unk>").
Reaching the operating point (recommended)
Plain beam search copies (~26% novelty). To reach the human point, use one of:
- Copy-penalty decoding (cheap, single pass): subtract a constant from the logits of tokens that appear in the source.
- Operating-point reranking: sample K candidates and pick the most novel one whose semantic preservation (LaBSE/GATE vs. source) stays above a floor.
Reference implementations are released with the paper code.
Results (official blind test set, 2,000 sentences)
In-band = % of outputs within the human band (novelty 40–70%, preservation ≥85).
| System | In-band % | Novelty | Preservation |
|---|---|---|---|
| Human (gold) | — | 56.1 | 86.7 |
| AraScholar + operating-point reranking | 76.6 | 55.7 | 89.3 |
| AraScholar + copy-penalty | 55.9 | 58.2 | 87.0 |
| GPT-4o | 49.4 | 65.1 | 86.4 |
| Claude Sonnet 4.6 | 18.1 | 75.4 | 82.5 |
| AraScholar (plain beam) | 10.0 | 25.6 | 93.9 |
A ~300M open model with decoding-time control matches or exceeds frontier LLMs on calibration for this task.
Intended use & limitations
For research and legitimate academic writing support (clarity, register). It operates at the sentence level. Like any paraphraser it is dual-use; deploy with attribution/integrity safeguards. Outputs should be checked for rare hallucinations, especially under high-temperature sampling.
Citation
@inproceedings{arascholar,
title = {Reaching the Human Operating Point: Controlled Academic Paraphrasing in Modern Standard Arabic},
author = {Anonymous},
year = {2025}
}